Sarvam Prism: an open-weight decision engine that answers in 44ms

Apache-2.0 weights on Hugging Face put routing, agent oversight and confidence gates in one forward pass — plus OpenAI's 948-exploit Linux kernel run.

Nowline NOV 8 9:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Decisions by scoring, not token generation

    Sarvam Prism is a 105B model (LoRA fine-tune, FP8) that scores a set of options in one forward pass instead of generating text — ~44ms median on a single B200, 131K-token context, Apache-2.0 weights on Hugging Face plus a hosted API. Self-host it as an intent/model router, an agent-run judge, or a confidence gate that acts only above a threshold: a fast, private decision layer with no per-token cost.

  • OpenAI's agent wrote 948 Linux kernel exploits

    An unreleased OpenAI cyber model flagged 2,308 kernel bugs, reproduced 1,724, and produced 948 working privilege-escalation exploits in a ~30-hour run — $1.2M in API credits, ~$1,279 per exploit. The model isn't public, but the demo is the clearest signal yet that automated vuln discovery is here; if you ship anything with native deps, assume attackers will soon run the same five-stage playbook.